fix: two manifest validation gaps (single-list NaN handling, RowFilter operator)

faircode/manifest.py

@@ -29,6 +29,17 @@ class RowFilter:
     not_equals: object | None = None
     notna: bool = False
 
+    def __post_init__(self):
+        # MANIFEST_SPEC.md requires "exactly one of the operators below" for a
+        # row_filters entry. A block with only `column` set (a blanked or
+        # copy-pasted YAML entry) would otherwise silently keep every row.
+        if (self.isin is None and self.not_isin is None and self.equals is None
+                and self.not_equals is None and not self.notna):
+            raise ValueError(
+                f"{self.column}: row filter needs one of "
+                f"isin / not_isin / equals / not_equals / notna"
+            )
+
     def apply(self, df: pd.DataFrame) -> pd.DataFrame:
         mask = pd.Series(True, index=df.index)
         if self.isin is not None:
@@ -121,11 +132,16 @@ def disadvantaged_mask(self, df: pd.DataFrame) -> tuple[pd.Series, pd.Series]:
                 known = col.isin(self.disadvantaged_values) | col.isin(self.advantaged_values)
                 disadv = col.isin(self.disadvantaged_values)
             elif self.disadvantaged_values is not None:
-                known = pd.Series(True, index=df.index)
+                # Only one list given: a NaN can't be classified either way
+                # (it matches neither isin() nor its complement), so exclude
+                # it via known_mask - matching the both-lists branch above
+                # and every other attribute type below, instead of silently
+                # routing it to the advantaged side.
+                known = col.notna()
                 disadv = col.isin(self.disadvantaged_values)
             elif self.advantaged_values is not None:
-                known = pd.Series(True, index=df.index)
-                disadv = ~col.isin(self.advantaged_values)
+                known = col.notna()
+                disadv = ~col.isin(self.advantaged_values) & known
             else:
                 raise ValueError(f"{self.name}: need disadvantaged_values or advantaged_values")
             return disadv, known

tests/test_manifest.py

@@ -202,6 +202,35 @@ def test_protected_attribute_categorical_complement():
     assert known.all()
 
 
+@pytest.mark.parametrize("kwargs", [
+    {"advantaged_values": ["white"]},
+    {"disadvantaged_values": ["black"]},
+])
+def test_categorical_single_list_excludes_nan_via_known_mask(kwargs):
+    # A NaN can't be classified when only one list is given, so known_mask
+    # must be False for it - previously the single-list branches hardcoded
+    # known=True and silently routed NaN to whichever side isin() landed it
+    # on (#547).
+    pa = ProtectedAttribute(name="g", type="categorical", column="g", **kwargs)
+    df = pd.DataFrame({"g": ["white", "black", None]})
+
+    disadv, known = pa.disadvantaged_mask(df)
+
+    assert known.tolist() == [True, True, False]
+    assert not bool(disadv.iloc[2])   # the NaN row is not counted as disadvantaged
+
+
+def test_row_filter_needs_at_least_one_operator():
+    # MANIFEST_SPEC.md requires exactly one operator; a bare column silently
+    # kept every row instead of raising (#549).
+    with pytest.raises(ValueError, match="row filter needs one of"):
+        RowFilter(column="race")
+
+    # notna=True counts as an operator; so does any of the others.
+    assert RowFilter(column="race", notna=True).column == "race"
+    assert RowFilter(column="race", equals="X").column == "race"
+
+
 def test_protected_attribute_numeric_threshold():
     pa = ProtectedAttribute(name="age", type="numeric_threshold", column="age",
                             threshold=30, disadvantaged="below")